llama.cpp not working properly on linux

[jylenhoftechsys]

Was trying to install with github backend llama.cpp with something like mise use -g github:ggml-org/llama.cpp... But install is incomplete because of the dynamic libraries present in the install directory which are not in the LD_LIBRARY_PATH

Was happy to see that someone has done an entry registry two days ago... but same problem with it

$ llama-cli --version
llama-cli: error while loading shared libraries: libllama-common.so.0: cannot open shared object file: No such file or directory
$

$ ls ~/.local/share/mise/installs/llama-cpp/8916/
export-graph-ops             libmtmd.so                     llama-gemma3-cli     llama-q8dot               test-backend-sampler         test-mtmd-c-api
libggml-base.so              libmtmd.so.0                   llama-gen-docs       llama-quantize            test-barrier                 test-opt
libggml-base.so.0            libmtmd.so.0.0.8916            llama-gguf           llama-qwen2vl-cli         test-c                       test-peg-parser
libggml-base.so.0.10.0       LICENSE                        llama-gguf-hash      llama-results             test-chat                    test-quantize-fns
libggml-cpu.so               llama-batched                  llama-gguf-split     llama-retrieval           test-chat-auto-parser        test-quantize-perf
libggml-cpu.so.0             llama-batched-bench            llama-idle           llama-save-load-state     test-chat-peg-parser         test-quantize-stats
libggml-cpu.so.0.10.0        llama-bench                    llama-imatrix        llama-server              test-chat-template           test-quant-type-selection
libggml-openvino.so          llama-cli                      llama-llava-cli      llama-simple              test-gbnf-validator          test-reasoning-budget
libggml-openvino.so.0        llama-completion               llama-lookahead      llama-simple-chat         test-gguf                    test-regex-partial
libggml-openvino.so.0.10.0   llama-convert-llama2c-to-ggml  llama-lookup         llama-speculative         test-gguf-model-data         test-rope
libggml.so                   llama-cvector-generator        llama-lookup-create  llama-speculative-simple  test-grammar-integration     test-sampling
libggml.so.0                 llama-debug                    llama-lookup-merge   llama-template-analysis   test-grammar-parser          test-state-restore-fragmented
libggml.so.0.10.0            llama-debug-template-parser    llama-lookup-stats   llama-tokenize            test-jinja                   test-thread-safety
libllama-common.so           llama-diffusion-cli            llama-minicpmv-cli   llama-tts                 test-json-partial            test-tokenizer-0
libllama-common.so.0         llama-embedding                llama-mtmd-cli       llama-vdot                test-json-schema-to-grammar  test-tokenizer-1-bpe
libllama-common.so.0.0.8916  llama-eval-callback            llama-mtmd-debug     test-alloc                test-llama-archs             test-tokenizer-1-spm
libllama.so                  llama-export-lora              llama-parallel       test-arg-parser           test-llama-grammar
libllama.so.0                llama-finetune                 llama-passkey        test-autorelease          test-log
libllama.so.0.0.8916         llama-fit-params               llama-perplexity     test-backend-ops          test-model-load-cancel
$

What is the best backend to use to be able to add some dynamic LD_LIBRARY_PATH addition when running binaries in this folder ? (seems to me that github and aqua doesn't have such option)

[jdx]

I think mise is probably the wrong tool for managing LD_LIBRARY_PATH at least currently

[jylenhoftechsys]

It's not directly managed we want... Just add on execution (and forget) the mise install dir to LD_LIBRARY_PATH, I think this is something which can be done easily in asdf backend, but I was asking because that's not a one you recommend much

[jylenhof]

Only tested on my workstation on linux, but my version seems to work a little bit better :
https://github.com/jylenhof/asdf-plugin-llama-cpp
That should works with macos also... if not some minors changed need to be done

If there's no better backend that this type of solution, maybe we need to change the actual registry entry to point to this version for MacOs and Linux, and still use github for windows....